app.py

import torch
from transformers import AutoFeatureExtractor, AutoModelForImageClassification
from einops import rearrange
import gradio
import call_labels

# define the feature extractor
extractor = AutoFeatureExtractor.from_pretrained("vincentclaes/mit-indoor-scenes")

# define the pretrained model
model = AutoModelForImageClassification.from_pretrained("vincentclaes/mit-indoor-scenes")

# retrieve the labels provided from MIT Indoor Scenes dataset (https://www.kaggle.com/itsahmad/indoor-scenes-cvpr-2019)
labels = call_labels.call_labels()

# call model.eval() to assert that we are evaluating the model and not updating the weights
model.eval()

# define the function used for model inference
def classify(image):
  # disable gradient calculation
  with torch.no_grad():
    # extract features from the image input
    inputs = extractor(images=image, return_tensors='pt')
    # call the logits parameter only (object: SequenceClassifierOutput)
    outputs = model(**inputs).logits
    # remove the batch dimension
    outputs = rearrange(outputs, '1 j->j')
    # use the softmax function to convert the logits into probabilities
    outputs = torch.nn.functional.softmax(outputs)
    # convert the data type from tensor to a numpy array
    outputs = outputs.cpu().numpy()
    # returns a key-value pair of the id labels and its corresponding probabilities
    return {labels[str(i)]: float(outputs[i]) for i in range(len(labels))}

# define the gradio interface
gradio.Interface(fn=classify,
                 inputs=gradio.inputs.Image(shape=(224,224),
                                            image_mode='RGB',
                                            source='upload',
                                            tool='editor',
                                            type='pil',
                                            label=None,
                                            optional=False),
                 outputs=gradio.outputs.Label(num_top_classes=5,
                                              type='auto'),
                 theme='grass',
                 examples=[['bedroom.jpg'],
                           ['bathroom_AS.jpg'],
                           ['samsung_room.jpg']],
                 live=True,
                 layout='horizontal',
                 title='Indoor Scene Recognition',
                 description='A smart and easy-to-use indoor scene classifier. Start by uploading an input image. The outputs are the top five indoor scene classes that best fit your input image.',
                 interpretation='default',
                 article='''<h2><b>Additional Information</b></h2><p style='text-align: justify'>This indoor scene classifier employs the <b><a href='https://huggingface.co/google/vit-base-patch16-224-in21k' target='_blank'>google/vit-base-patch16-224-in21k</a></b>, a <b>Visual Transformer (ViT)</b> model pre-trained on ImageNet-21k (14 million images, 21,843 classes) at resolution 224x224 and was first introduced in the paper <b><a href='https://arxiv.org/abs/2010.11929' target='_blank'>An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale</a></b> by Dosovitskiy et al. The original GitHub repository of the Visual Transformer is found in <b><a href='https://github.com/google-research/vision_transformer' target='_blank'>this link</a></b>. This Visual Transformer model was fine-tuned on the <b><a href='https://www.kaggle.com/itsahmad/indoor-scenes-cvpr-2019' target='_blank'>MIT Indoor Scenes</a></b> from Kaggle. The source model from Hugging Face is found in <b><a href='https://huggingface.co/vincentclaes/mit-indoor-scenes' target='_blank'>this link</a></b>.</p>''',
                 allow_flagging='never').launch()